0
$\begingroup$

Is there way to segment an image without labeling/classification, as well as supervised learning?

For an illustrative example, if one considers an image with a dog and a cup (we don't particularly care what is in the picture), what are viable approaches (if there exist any) to segment these objects in order to perform localization afterwards?

How can one detect arbitrary objects in an image?

So far, I tried a level set approach using a Hamilton-Jacobi Equation to model the energy within the picture. However, this is extremely slow (since iterative) and highly depends on the preprocessing of the image to detect anything.

$\endgroup$
2
  • 1
    $\begingroup$ Are you looking for an unsupervised learning technique to find the objects? You could cluster the image (e.g. with k-means) - if you're lucky, the clusters will be the objects. Maybe there's some specific model/algorithm that was specifically created to do your task. $\endgroup$
    – nbro
    Apr 27 at 10:22
  • $\begingroup$ Yes some kind of unsupervised approach where no knowledge of the object is a priori known. Thanks for the idea I will have a look into clustering. Ultimately, my goal is to compute the positions of these objects robustly. (And maybe extend it to tracking using MPC, to extend the research) $\endgroup$
    – Astraeus
    Apr 27 at 11:03

1 Answer 1

1
$\begingroup$

You could run your images through a pretrained CNN for feature extraction, then run an unsupervised algorithm to see if there are any obvious groups of "dogs" and "cups".

Don't know how you would procced with localization though. I'm afraid you'll have to hand-label your images via some tool. Maybe there are some pretrained models for localization you could use.

$\endgroup$
1
  • $\begingroup$ That is kinda what I hoped to avoid, to have some pretrained CNNs, since it seems (to me) to restricting with respect to the types of objects I can detect. I think it would be interesting in general to be able to get for example the outlines of an arbitrary shape. With the Hamilton-Jacobi Level-Set approach I am able to do so to some extent but it is quite slow, so I cannot transfer it to videos. The suggestion of @nbro with k-means at least sped up this approach. However still not nearly as fast as a pretrained CNN. Anyway still thanks for your suggestion ! $\endgroup$
    – Astraeus
    May 2 at 8:42

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .